Llama 3.2 90B Vision Instruct

Meta's Llama 3.2 90B Vision Instruct is a 90-billion-parameter vision-language model released in September 2024, extending the Llama 3.1 architecture with a dedicated vision encoder under the Llama 3 community license. It is Meta's first openly licensed multimodal Llama, targeting image understanding tasks — ChartQA, DocVQA, and OCR-heavy document workflows where you need a single model to handle both structured text and visual inputs. On ChartQA it approaches or matches proprietary mid-tier VLMs. At 90B parameters the hosted inference cost sits above the 70B text tier; plan accordingly if you're running high-volume image pipelines. The 131K context window holds, so long documents with embedded images are tractable. Expect broader provider coverage and lower per-token pricing as the 3.2 multimodal lineup matures.